Internal Di usion Limited Aggregationsava/other_files/idla_siegen_2011.pdf · Internal Di usion...

Graz University of Technology

Warm-up: DLA Internal DLA References

Internal Diffusion Limited Aggregation(joint work with Wilfried Huss)

Ecaterina Sava

May 24, 2011Miniworkshop, Siegen University

Outline

1 Warm-up: DLA

2 Internal DLA

3 References

Introduced in physics by Sander and Witten [’81], as a model offractal growth. The growth rule is extremly simple:

start with only the origin of some coordinate system, which isoccupied.

repeatedly send random walks “in from ∞“.

each walk stops when it neighbors the previously occupiedcluster; the set of occupied sites is called the DLA cluster.

Question: how does the aggregation cluster DLA obtained inthis way, look like?

DLA has a shape that is widely belived to have fractalcharacteristics.

DLA tends to build irregularities.

DLA 33000 particles, center initially occupied

Different colors = different arrival times of the random walkers.

Random walkers sticking to a straight line

Internal DLARandom growth model, internal version of DLA, which contraryto DLA, tends to eliminate irregularities.

Let G be some infinite graph, and o some fixed vertex.one by one, particles perform discrete-time random walks.each particle starts from o and moves until it reaches a siteunoccupied previously, where it stops.get a random subset of n occupied sites in G : internal DLAcluster A(n) → the resulting random cluster of occupied sitesafter the nth particle stops.

Growth rule: Let A(0) = {o} and define

A(n + 1) = A(n) ∪ {X n(τn)},where X 1,X 2, . . . are independent random walks starting at o,and

τn = min{t : X n(t) /∈ A(n − 1)}.

Main question: limiting shape of A(n) as n→∞?

Let G be some infinite graph, and o some fixed vertex.one by one, particles perform discrete-time random walks.each particle starts from o and moves until it reaches a siteunoccupied previously, where it stops.get a random subset of n occupied sites in G : internal DLAcluster A(n) → the resulting random cluster of occupied sitesafter the nth particle stops.Growth rule: Let A(0) = {o} and define

τn = min{t : X n(t) /∈ A(n − 1)}.

Let G be some infinite graph, and o some fixed vertex.one by one, particles perform discrete-time random walks.each particle starts from o and moves until it reaches a siteunoccupied previously, where it stops.get a random subset of n occupied sites in G : internal DLAcluster A(n) → the resulting random cluster of occupied sitesafter the nth particle stops.Growth rule: Let A(0) = {o} and define

τn = min{t : X n(t) /∈ A(n − 1)}.

DLA and internal DLA : comparison

DLA: tendrils result from the fact that the particles tend to hitfirst the neighbourhood of extreme sites in the occupied cluster ⇒fractal structure.internal DLA: particles diffusing through the interior of theoccupied cluster are most likely to stop at unoccupied sites thatare closest to 0 ⇒ A(n) tends to eliminate irregularities ⇒expected to grow like an expanding ball on a regular graph.

Theorem (Lawler-Bramson-Griffeath ’92)

For simple random walks on Zd , d ≥ 2, the limiting shape ofinternal DLA is a ball: ∀ε > 0, with probability 1:

Br(1−ε) ⊂ A(πr2) ⊂ Br(1+ε), eventually,

Question: what about fluctuations for internal DLA, i.e. howsmooth the surface formed by internal DLA can be?

The cluster A(n), n=100000

Fluctuations on Zd : history

Lawler [’95]: with probability 1

Br−r1/3 log2 r ⊂ A(πr2) ⊂ Br+r1/3 log4 r

Can the errors be of order o(nα), for α < 1/3? Indeed thereare only logarithmic fluctuations.

Jerison-Levine-Sheffield [’10]: with probability 1

Br−C log r ⊂ A(πr2) ⊂ Br+C log r , eventually.

Asselah-Gaudilliere [’10]: independently obtained

Br−C log r ⊂ A(πr2) ⊂ Br+C log2 r , eventually.

For d ≥ 3: Jerison-Levine-Sheffield [’10] andAsselah-Gaudilliere [’10]

Br−C√log r ⊂ A(ωd rd) ⊂ B

r+C√

log2 r, eventually,

for a constant C depending only on d . ωd is the volume ofthe d-dimensional Euclidean ball of radius 1.

r+C√

log2 r, eventually,

r+C√

log2 r, eventually,

r+C√

log2 r, eventually,

IDLA on different state spaces

Green function: G (x , y) = Ex

[#{t ≥ 0 : X (t) = y}

Levelsets of the Green function: {x ∈ G : G (o, x) ≥ N}.

TheoremThe levelsets of the Green function are the limiting shape for IDLAwith probability 1, for:

(Lawler-Bramson-Griffeath ’92) simple random walk on Zd .

(Blachere, ’02) symmetric random walks on Zd .

(Blachere-Brofferio ’06) symmetric random walks on Cayleygraphs of finitely generated groups with exponential growth.

(Huss ’07) strongly reversible, uniformly irreducible randomwalks on non-amenable graphs.

[#{t ≥ 0 : X (t) = y}

Does this theorem hold for IDLA in general?

Counter examples:

Random walk with drift in Z2

Simple random walk on the comb

Does this theorem hold for IDLA in general? NO

Counter examples:

Does this theorem hold for IDLA in general? NO

Counter examples:

Internal DLA on the comb

Comb C2 is the graph obtained fromZ2 by deleting all horizontal edges,except for x-axis.

consider simple random walk on C2.

p(x , y) =1

d(x), for all x ∈ C2,

where d(x) is the degree of x .

Perform internal DLA with n simple random walks starting at theorigin o = (0, 0) ∈ C2. We obtain the internal DLA cluster A(n),random subset of C2 with n elements.

What is the limiting shape A(n)?

p(x , y) =1

A(n) for n = 500 and n = 1000.

the set in the figure grows like n2/3 inthe vertical direction and like n1/3 inthe horizontal direction.

want to prove that this is the limitingshape of internal DLA on the comb C2.

unfortunatelly, up to now we can proveonly an inner bound: with probability 1

Bn(1−ε) ⊂ A(n),

{(x , y) ∈ C2 :

( |y |l

≤ n1/3

}Outer bound: someone in the audience[May ’11]: A(n) ⊂ Bn(1+ε) ?

A(n) for n = 500 and n = 1000.

Bn(1−ε) ⊂ A(n),

{(x , y) ∈ C2 :

( |y |l

≤ n1/3

A(n) for n = 500 and n = 1000.

Bn(1−ε) ⊂ A(n),

{(x , y) ∈ C2 :

( |y |l

≤ n1/3

A(n) for n = 500 and n = 1000.

Bn(1−ε) ⊂ A(n),

{(x , y) ∈ C2 :

( |y |l

≤ n1/3

Theorem (Huss - S. ’10)

Let A(n) be the internal DLA cluster after n random walks start atthe origin of C2. Then, for all ε > 0, we have with probability 1

Bn(1−ε) ⊂ A(n), for all sufficiently large n.

Proof sketch.

Inspired by the Lawler-Bramson-Griffeath argument.

By Borel-Cantelli Lemma, a sufficient condition for provingthe inner bound is∑

n≥n0

∑z∈Bn(1−ε)

P[z /∈ An] <∞.

Fix z ∈ Bn. We want an upper bound for P[z /∈ A(n)

n random walks start at o ⇒ A(n).

M = # of walks that visit z beforeleaving Bn,

L = # of walks that visit z afterleaving A(i), while still in Bn,1 ≤ i ≤ n.

If L < M then z ∈ A(n) and

{z /∈ A(n)} ⊂ {M = L}.

L and M are sums of indicator rv’s.

the summands of L are dependent.

bound L by a sum of i.i.d rv’s

only the walks that leave A(i) inBn contribute to L: start one newwalk from every point in Bn wherethe cluster is left.

{z /∈ A(n)} ⊂ {M = L}.

Proof sketch

enlarge the index set to all of Bn

L = # of new walks that hit z before leaving Bn. Then

L ≤ L,

andP[z /∈ A(n)] ≤ P[M = L] ≤ P[M ≤ L].

we show that∑n≥nε

∑z∈Bn(1−ε)

P[M ≤ L] ≤ 4∑n≥nε

n exp{−Cεn2/3} <∞,

which proves the inner bound

P[Bn(1−ε) ⊂ An, for all n ≥ nε

Proof sketch

enlarge the index set to all of BnL = # of new walks that hit z before leaving Bn. Then

L ≤ L,

andP[z /∈ A(n)] ≤ P[M = L] ≤ P[M ≤ L].

∑z∈Bn(1−ε)

P[M ≤ L] ≤ 4∑n≥nε

n exp{−Cεn2/3} <∞,

Proof sketch

enlarge the index set to all of BnL = # of new walks that hit z before leaving Bn. Then

L ≤ L,

andP[z /∈ A(n)] ≤ P[M = L] ≤ P[M ≤ L].

∑z∈Bn(1−ε)

P[M ≤ L] ≤ 4∑n≥nε

n exp{−Cεn2/3} <∞,

Outlook

For internal DLA on the comb lattice C, an outer bound

A(n) ⊂ Bn(1+ε)

is still needed.

In all previous proofs, internal DLA clusters A(n) grow uniformly,and this makes easy the study of random walks. In our case, this isviolated, since the sets Bn grow with rate n1/3 in the x-directionand with rate n2/3 in the y -direction.The study of the harmonic measure and the Green functionstopped on sets Bn may help.

Outlook

A(n) ⊂ Bn(1+ε)

is still needed.In all previous proofs, internal DLA clusters A(n) grow uniformly,and this makes easy the study of random walks. In our case, this isviolated, since the sets Bn grow with rate n1/3 in the x-directionand with rate n2/3 in the y -direction.

The study of the harmonic measure and the Green functionstopped on sets Bn may help.

Outlook

A(n) ⊂ Bn(1+ε)

is still needed.In all previous proofs, internal DLA clusters A(n) grow uniformly,and this makes easy the study of random walks. In our case, this isviolated, since the sets Bn grow with rate n1/3 in the x-directionand with rate n2/3 in the y -direction.The study of the harmonic measure and the Green functionstopped on sets Bn may help.

Sierpinski carpetGraphical Sierpinski carpet in dimension 2: infinite graph derivedfrom the Sierpinski carpet - a fractal created from the unit squarein R2 by dividing it into 9 equal squares of which the one in thecenter is deleted. The same procedure is then repeated recursivelyto the remaining 8 squares.

Figure: IDLA clusters on the Sierpinski carpet for 10000 particles.

Internal DLA on Sierpinski carpet

Problems:

the internal DLA cluster does not seem to have an uniquescaling limit.

simulations sugest that may be a whole family of scalinglimits.

this scaling limits seem to have a fractal boundary.

Problems:

n = 275

n = 275 n = 2175

n = 275 n = 2175 n = 17500

n = 275 n = 2175 n = 17500 n = 140000

n = 200

n = 275 n = 2175 n = 17500 n = 140000

n = 200 n = 1500

n = 275 n = 2175 n = 17500 n = 140000

n = 200 n = 1500 n = 12500

n = 275 n = 2175 n = 17500 n = 140000

n = 200 n = 1500 n = 12500 n = 98000

References

G. Lawler, M. Bramson and D. Griffeath,Internal Diffusion Limited Aggregation, Ann. Probab. 20, no. 4(1992), 2117–2140.

L. Levine, and Y. Peres,Scaling Limits for Internal Aggregation Models with MultipleSources, Journal d’Analyse Mathematique 111 (2010)151–219.

W. Huss, and E. Sava,Internal Aggregation Models on the Comb Lattice, preprint,2011.

Thank your for your interest!

Internal Di usion Limited Aggregationsava/other_files/idla_siegen_2011.pdf · Internal Di usion...

Documents

Transcript of Internal Di usion Limited Aggregationsava/other_files/idla_siegen_2011.pdf · Internal Di usion...